The most common phrase in medical AI marketing is 'FDA authorized.' Yet regulatory authorization and proven clinical effectiveness are two distinct things: the former indicates that a device meets applicable premarket requirements, while the latter requires well-designed clinical studies demonstrating improvement in target endpoints. Four studies published between 2025 and 2026 have, for the first time, quantified this gap—revealing it to be larger than most imagined in the U.S. market, while the regulatory system has only just begun building tools to address it.
## How Large Is the Gap? Three Independent Counts, Same Direction
The first is a systematic review published in JAMA Network Open in 2025, covering all FDA premarket authorizations from November 1995 through June 2024: a total of 950 AI/ML medical devices, with radiology accounting for 723 (76%); 97% (924 devices) were cleared via the 510(k) pathway—that is, by demonstrating substantial equivalence to a legally marketed 'predicate device,' a route that does not require submission of independent clinical data. Researchers coded test types for 717 radiology devices with available submission documents, yielding three striking figures: only 33 (5%) underwent prospective testing; only 56 (8%) included human-in-the-loop testing; and only 208 (29%) included clinical testing. Just 15 devices met both prospective and clinical testing criteria, and only six satisfied all three.
The growth curve is equally steep: 33 devices were authorized across the 20 years from 1995 to 2015, whereas 221 were authorized in 2023 alone.
The second count, published in the Journal of Medical Artificial Intelligence in 2025 and updated through May 2025, identified 1,016 FDA-authorized AI/ML devices, of which 38.8% were authorized after January 2023; 96.5% used the 510(k) pathway, De Novo accounted for 3%, and the most stringent PMA pathway accounted for less than 1%; radiology comprised 76.1%, and cardiovascular devices 9.8%. Transparency dimensions are even more notable: only 55.8% disclosed algorithmic details beyond generic descriptions in their premarket summaries, and fewer than 20% cited prospective validation data.
The third source is the FDA’s own official list page. This list explicitly states it is 'not a comprehensive resource for AI medical devices' and identifies devices primarily based on AI-related terminology in authorization document summaries. The FDA also notes it is exploring ways to flag devices incorporating foundation models—including large language models and multimodal architectures. In other words, even the question 'How many AI devices exist?' yields only an approximate answer from regulators’ public statements.
## Authorization Is Not the Endpoint—But Postmarket Surveillance Looks Like It Is
What happens post-authorization? A 2026 study in npj Digital Medicine tracked 956 radiology AI devices authorized between September 1995 and September 2025: 429 (44.87%) underwent software version updates, involving 94 manufacturers and 163 products; adverse events linked to software defects involved 15 products, of which only four (26.67%) subsequently issued version updates, and only three (20%) initiated voluntary recalls; there were 124 recall reports citing software defects (68.13% of all recall reports), and 38 of these were resolved via software updates after the recall. The authors’ core conclusion: the vast majority of manufacturers lack a closed-loop risk management system covering 'adverse event → recall → software update'—the three stages operate in isolation, making defect identification difficult and limiting containment of risk propagation.
## The First Repair Tools—and Their Blind Spots
Regulatory responses have already emerged. In December 2024, the FDA issued its final guidance on 'Predetermined Change Control Plans' (PCCPs), permitting manufacturers to pre-specify how algorithms may change—e.g., via retraining, compatibility expansion, or performance optimization—without requiring new submissions for each modification. On January 7, 2025, the FDA released a draft guidance on lifecycle management of software functions for AI devices. A 2026 scoping review in Radiology assessed real-world adoption of these tools: as of April 2026, 170 devices had received authorization with PCCPs, including 34 in radiology; of those 34, 22 (65%) were authorized within 2025—the year following the December 2024 final guidance—clearly indicating guideline-driven uptake.
Yet transparency remains concerning. Using an 8-point scoring rubric (with two independent reviewers), PCCP documentation scores ranged from 0 to 8, with a mean of only 5. Continuous performance monitoring and pre-specified retraining drift triggers are commonly omitted from publicly available summaries. Among the 34 radiology PCCP devices, nine (27%) exhibited inconsistencies between FDA’s public database and individual summary documents, requiring manual adjudication. In plain terms: the mechanism allowing manufacturers to 'pre-declare how they will change' is now in place—but the public and hospitals cannot see whether 'those changes are under active monitoring.'
## Counterpoint: Transparency ≠ Rigor
A position paper in Radiology: Artificial Intelligence offers a necessary methodological rebuttal to these counts: one must distinguish between validation rigor (quality and comprehensiveness of evidence) and validation transparency (public accessibility of evidence)—performance data volume in FDA submissions and public transparency exhibit an inverse relationship, meaning abbreviated public summaries do not necessarily indicate absent or poor validation. This rebuttal is valid, so it is included here: the '5% prospective testing' metric reflects only codable evidence in publicly accessible documents; the true distribution of validation strength may be higher—or lower—yet external parties have no way to know. That itself is the problem: verification that patients and purchasers cannot independently verify effectively does not exist in the realm of public evidence.
Two additional factual points merit juxtaposition. First, the discussion section of the JAMA Network Open review notes that after tightening regulation, the EU saw a decline in approvals of radiology AI devices alongside a rise in peer-reviewed publications reporting clinical outcomes—tightening did not stifle innovation but redirected it toward better-evidenced development. Second, when radiologists use AI-assisted image interpretation, high-performing clinicians maintain performance levels, while low-performing ones do not necessarily improve—'authorization + deployment' does not equal 'clinical benefit,' and human heterogeneity persists.
## Three Takeaways You Can Use
First, for hospitals and procurement teams: demote 'FDA authorization' from a conclusion to metadata. During due diligence, ask at least five questions verifiable from public materials—Was validation design retrospective, prospective, or human-in-the-loop? Does the device have a PCCP, and how does its public documentation score? What is its history of postmarket software updates and recalls? Are performance monitoring and drift-triggered retraining disclosed? What is the demographic composition of its training data?
Second, for regulatory observers: PCCP adoption driven by guidance is encouraging, but the asymmetry between 'publicly declared change plans' and 'non-public monitoring data' must be addressed in the next round of guidance.
Third, for general readers: the next time you encounter medical AI marketing touting 'FDA authorization,' remember that on the same device, 'FDA authorized,' 'prospectively clinically validated,' and 'supported by closed-loop postmarket surveillance' are three logically independent propositions. Compressing them into a single intuitive notion of 'official certification' is precisely how this gap replicates at the cognitive level.
A critical question worth debating: if the 'substantial equivalence' logic underlying the 510(k) pathway predates modern software iteration, can an algorithm retrained annually still meaningfully be considered the 'same device' as its predicate?
Source: FDA Approval of Artificial Intelligence and Machine Learning Devices in Radiology: A Systematic Review. JAMA Network Open 2025. https://doi.org/10.1001/jamanetworkopen.2025.42338